mirror of
https://github.com/belt-sh/cli.git
synced 2026-09-14 16:31:15 +08:00
7fc61fd2c2
docs, install scripts, agent context (CLAUDE.md, AGENTS.md), examples, and issue templates. source is closed — this repo serves as the public resource hub for belt cli usage.
1.8 KiB
1.8 KiB
knowledge management
belt's knowledge system gives agents persistent memory across sessions. save discoveries, patterns, references — anything worth remembering.
types
| type | use for |
|---|---|
observation |
things you noticed or learned |
concept |
mental models, explanations |
skill |
reusable workflows with steps |
reference |
links to external resources |
preference |
user preferences, style choices |
saving knowledge
# from stdin (quick observations)
echo "react server components can't use useState" | belt know create - --name rsc-no-state --type observation
# from a file
belt know create ./api-patterns.md --type concept
# from a directory (skills with multiple files)
belt know create ./my-workflow/ --type skill
# with explicit metadata
belt know create ./notes.md --name deployment-checklist --type reference
searching
# semantic search
belt know search "error handling patterns"
belt know search "react hooks"
# list by type
belt know list --type skill
belt know list --type observation
# get full details
belt know get myteam/deployment-checklist
agent usage
agents can use knowledge to build context:
# before starting work, check what's known
belt know search "auth middleware" --json
# after discovering something, save it
belt know create - --name finding --type observation <<EOF
the auth middleware caches tokens for 5 minutes.
invalidation requires a redis pub/sub message.
EOF
knowledge in the claude code plugin
with the belt-sh/skills plugin installed, knowledge is automatically:
- searched on every prompt (via the suggest hook)
- captured from sessions (via the stop hook with haiku evaluation)
- deduplicated before saving (via the knowledge-curator agent)